Unserious surprise summarised it nicely for me; It does not mean that Markov chains are better - something trained to make predictions should ideally not fall

Unserious surprise summarised it nicely for me; It does not mean that Markov chains are better - something trained to make predictions should ideally not fall too far away from our own internal prediction engines (which have been honed across aeons). It's that it starts to come close that's the problem (or cause); it's the uncanny valley for text.